1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2013 by Delphix. All rights reserved.
24 * Copyright (c) 2013 Steven Hartland. All rights reserved.
25 */
26
27 #include <sys/dsl_pool.h>
28 #include <sys/dsl_dataset.h>
29 #include <sys/dsl_prop.h>
30 #include <sys/dsl_dir.h>
31 #include <sys/dsl_synctask.h>
32 #include <sys/dsl_scan.h>
33 #include <sys/dnode.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/arc.h>
37 #include <sys/zap.h>
38 #include <sys/zio.h>
39 #include <sys/zfs_context.h>
40 #include <sys/fs/zfs.h>
41 #include <sys/zfs_znode.h>
42 #include <sys/spa_impl.h>
43 #include <sys/dsl_deadlist.h>
44 #include <sys/bptree.h>
45 #include <sys/zfeature.h>
46 #include <sys/zil_impl.h>
47 #include <sys/dsl_userhold.h>
48
49 int zfs_no_write_throttle = 0;
50 int zfs_write_limit_shift = 3; /* 1/8th of physical memory */
51 int zfs_txg_synctime_ms = 1000; /* target millisecs to sync a txg */
52
53 uint64_t zfs_write_limit_min = 32 << 20; /* min write limit is 32MB */
54 uint64_t zfs_write_limit_max = 0; /* max data payload per txg */
55 uint64_t zfs_write_limit_inflated = 0;
56 uint64_t zfs_write_limit_override = 0;
57
58 kmutex_t zfs_write_limit_lock;
59
60 static pgcnt_t old_physmem = 0;
61
62 hrtime_t zfs_throttle_delay = MSEC2NSEC(10);
63 hrtime_t zfs_throttle_resolution = MSEC2NSEC(10);
64
65 int
66 dsl_pool_open_special_dir(dsl_pool_t *dp, const char *name, dsl_dir_t **ddp)
67 {
68 uint64_t obj;
69 int err;
70
71 err = zap_lookup(dp->dp_meta_objset,
72 dp->dp_root_dir->dd_phys->dd_child_dir_zapobj,
73 name, sizeof (obj), 1, &obj);
74 if (err)
75 return (err);
76
77 return (dsl_dir_hold_obj(dp, obj, name, dp, ddp));
78 }
79
80 static dsl_pool_t *
81 dsl_pool_open_impl(spa_t *spa, uint64_t txg)
82 {
83 dsl_pool_t *dp;
84 blkptr_t *bp = spa_get_rootblkptr(spa);
85
86 dp = kmem_zalloc(sizeof (dsl_pool_t), KM_SLEEP);
87 dp->dp_spa = spa;
88 dp->dp_meta_rootbp = *bp;
89 rrw_init(&dp->dp_config_rwlock, B_TRUE);
90 dp->dp_write_limit = zfs_write_limit_min;
91 txg_init(dp, txg);
92
93 txg_list_create(&dp->dp_dirty_datasets,
94 offsetof(dsl_dataset_t, ds_dirty_link));
95 txg_list_create(&dp->dp_dirty_zilogs,
96 offsetof(zilog_t, zl_dirty_link));
97 txg_list_create(&dp->dp_dirty_dirs,
98 offsetof(dsl_dir_t, dd_dirty_link));
99 txg_list_create(&dp->dp_sync_tasks,
100 offsetof(dsl_sync_task_t, dst_node));
101
102 mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
103
104 dp->dp_vnrele_taskq = taskq_create("zfs_vn_rele_taskq", 1, minclsyspri,
105 1, 4, 0);
106
107 return (dp);
108 }
109
110 int
111 dsl_pool_init(spa_t *spa, uint64_t txg, dsl_pool_t **dpp)
112 {
113 int err;
114 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
115
116 err = dmu_objset_open_impl(spa, NULL, &dp->dp_meta_rootbp,
117 &dp->dp_meta_objset);
118 if (err != 0)
119 dsl_pool_close(dp);
120 else
121 *dpp = dp;
122
123 return (err);
124 }
125
126 int
127 dsl_pool_open(dsl_pool_t *dp)
128 {
129 int err;
130 dsl_dir_t *dd;
131 dsl_dataset_t *ds;
132 uint64_t obj;
133
134 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
135 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
136 DMU_POOL_ROOT_DATASET, sizeof (uint64_t), 1,
137 &dp->dp_root_dir_obj);
138 if (err)
139 goto out;
140
141 err = dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
142 NULL, dp, &dp->dp_root_dir);
143 if (err)
144 goto out;
145
146 err = dsl_pool_open_special_dir(dp, MOS_DIR_NAME, &dp->dp_mos_dir);
147 if (err)
148 goto out;
149
150 if (spa_version(dp->dp_spa) >= SPA_VERSION_ORIGIN) {
151 err = dsl_pool_open_special_dir(dp, ORIGIN_DIR_NAME, &dd);
152 if (err)
153 goto out;
154 err = dsl_dataset_hold_obj(dp, dd->dd_phys->dd_head_dataset_obj,
155 FTAG, &ds);
156 if (err == 0) {
157 err = dsl_dataset_hold_obj(dp,
158 ds->ds_phys->ds_prev_snap_obj, dp,
159 &dp->dp_origin_snap);
160 dsl_dataset_rele(ds, FTAG);
161 }
162 dsl_dir_rele(dd, dp);
163 if (err)
164 goto out;
165 }
166
167 if (spa_version(dp->dp_spa) >= SPA_VERSION_DEADLISTS) {
168 err = dsl_pool_open_special_dir(dp, FREE_DIR_NAME,
169 &dp->dp_free_dir);
170 if (err)
171 goto out;
172
173 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
174 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj);
175 if (err)
176 goto out;
177 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
178 dp->dp_meta_objset, obj));
179 }
180
181 if (spa_feature_is_active(dp->dp_spa,
182 &spa_feature_table[SPA_FEATURE_ASYNC_DESTROY])) {
183 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
184 DMU_POOL_BPTREE_OBJ, sizeof (uint64_t), 1,
185 &dp->dp_bptree_obj);
186 if (err != 0)
187 goto out;
188 }
189
190 if (spa_feature_is_active(dp->dp_spa,
191 &spa_feature_table[SPA_FEATURE_EMPTY_BPOBJ])) {
192 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
193 DMU_POOL_EMPTY_BPOBJ, sizeof (uint64_t), 1,
194 &dp->dp_empty_bpobj);
195 if (err != 0)
196 goto out;
197 }
198
199 err = zap_lookup(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
200 DMU_POOL_TMP_USERREFS, sizeof (uint64_t), 1,
201 &dp->dp_tmp_userrefs_obj);
202 if (err == ENOENT)
203 err = 0;
204 if (err)
205 goto out;
206
207 err = dsl_scan_init(dp, dp->dp_tx.tx_open_txg);
208
209 out:
210 rrw_exit(&dp->dp_config_rwlock, FTAG);
211 return (err);
212 }
213
214 void
215 dsl_pool_close(dsl_pool_t *dp)
216 {
217 /* drop our references from dsl_pool_open() */
218
219 /*
220 * Since we held the origin_snap from "syncing" context (which
221 * includes pool-opening context), it actually only got a "ref"
222 * and not a hold, so just drop that here.
223 */
224 if (dp->dp_origin_snap)
225 dsl_dataset_rele(dp->dp_origin_snap, dp);
226 if (dp->dp_mos_dir)
227 dsl_dir_rele(dp->dp_mos_dir, dp);
228 if (dp->dp_free_dir)
229 dsl_dir_rele(dp->dp_free_dir, dp);
230 if (dp->dp_root_dir)
231 dsl_dir_rele(dp->dp_root_dir, dp);
232
233 bpobj_close(&dp->dp_free_bpobj);
234
235 /* undo the dmu_objset_open_impl(mos) from dsl_pool_open() */
236 if (dp->dp_meta_objset)
237 dmu_objset_evict(dp->dp_meta_objset);
238
239 txg_list_destroy(&dp->dp_dirty_datasets);
240 txg_list_destroy(&dp->dp_dirty_zilogs);
241 txg_list_destroy(&dp->dp_sync_tasks);
242 txg_list_destroy(&dp->dp_dirty_dirs);
243
244 arc_flush(dp->dp_spa);
245 txg_fini(dp);
246 dsl_scan_fini(dp);
247 rrw_destroy(&dp->dp_config_rwlock);
248 mutex_destroy(&dp->dp_lock);
249 taskq_destroy(dp->dp_vnrele_taskq);
250 if (dp->dp_blkstats)
251 kmem_free(dp->dp_blkstats, sizeof (zfs_all_blkstats_t));
252 kmem_free(dp, sizeof (dsl_pool_t));
253 }
254
255 dsl_pool_t *
256 dsl_pool_create(spa_t *spa, nvlist_t *zplprops, uint64_t txg)
257 {
258 int err;
259 dsl_pool_t *dp = dsl_pool_open_impl(spa, txg);
260 dmu_tx_t *tx = dmu_tx_create_assigned(dp, txg);
261 objset_t *os;
262 dsl_dataset_t *ds;
263 uint64_t obj;
264
265 rrw_enter(&dp->dp_config_rwlock, RW_WRITER, FTAG);
266
267 /* create and open the MOS (meta-objset) */
268 dp->dp_meta_objset = dmu_objset_create_impl(spa,
269 NULL, &dp->dp_meta_rootbp, DMU_OST_META, tx);
270
271 /* create the pool directory */
272 err = zap_create_claim(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
273 DMU_OT_OBJECT_DIRECTORY, DMU_OT_NONE, 0, tx);
274 ASSERT0(err);
275
276 /* Initialize scan structures */
277 VERIFY0(dsl_scan_init(dp, txg));
278
279 /* create and open the root dir */
280 dp->dp_root_dir_obj = dsl_dir_create_sync(dp, NULL, NULL, tx);
281 VERIFY0(dsl_dir_hold_obj(dp, dp->dp_root_dir_obj,
282 NULL, dp, &dp->dp_root_dir));
283
284 /* create and open the meta-objset dir */
285 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, MOS_DIR_NAME, tx);
286 VERIFY0(dsl_pool_open_special_dir(dp,
287 MOS_DIR_NAME, &dp->dp_mos_dir));
288
289 if (spa_version(spa) >= SPA_VERSION_DEADLISTS) {
290 /* create and open the free dir */
291 (void) dsl_dir_create_sync(dp, dp->dp_root_dir,
292 FREE_DIR_NAME, tx);
293 VERIFY0(dsl_pool_open_special_dir(dp,
294 FREE_DIR_NAME, &dp->dp_free_dir));
295
296 /* create and open the free_bplist */
297 obj = bpobj_alloc(dp->dp_meta_objset, SPA_MAXBLOCKSIZE, tx);
298 VERIFY(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
299 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx) == 0);
300 VERIFY0(bpobj_open(&dp->dp_free_bpobj,
301 dp->dp_meta_objset, obj));
302 }
303
304 if (spa_version(spa) >= SPA_VERSION_DSL_SCRUB)
305 dsl_pool_create_origin(dp, tx);
306
307 /* create the root dataset */
308 obj = dsl_dataset_create_sync_dd(dp->dp_root_dir, NULL, 0, tx);
309
310 /* create the root objset */
311 VERIFY0(dsl_dataset_hold_obj(dp, obj, FTAG, &ds));
312 os = dmu_objset_create_impl(dp->dp_spa, ds,
313 dsl_dataset_get_blkptr(ds), DMU_OST_ZFS, tx);
314 #ifdef _KERNEL
315 zfs_create_fs(os, kcred, zplprops, tx);
316 #endif
317 dsl_dataset_rele(ds, FTAG);
318
319 dmu_tx_commit(tx);
320
321 rrw_exit(&dp->dp_config_rwlock, FTAG);
322
323 return (dp);
324 }
325
326 /*
327 * Account for the meta-objset space in its placeholder dsl_dir.
328 */
329 void
330 dsl_pool_mos_diduse_space(dsl_pool_t *dp,
331 int64_t used, int64_t comp, int64_t uncomp)
332 {
333 ASSERT3U(comp, ==, uncomp); /* it's all metadata */
334 mutex_enter(&dp->dp_lock);
335 dp->dp_mos_used_delta += used;
336 dp->dp_mos_compressed_delta += comp;
337 dp->dp_mos_uncompressed_delta += uncomp;
338 mutex_exit(&dp->dp_lock);
339 }
340
341 static int
342 deadlist_enqueue_cb(void *arg, const blkptr_t *bp, dmu_tx_t *tx)
343 {
344 dsl_deadlist_t *dl = arg;
345 dsl_deadlist_insert(dl, bp, tx);
346 return (0);
347 }
348
349 void
350 dsl_pool_sync(dsl_pool_t *dp, uint64_t txg)
351 {
352 zio_t *zio;
353 dmu_tx_t *tx;
354 dsl_dir_t *dd;
355 dsl_dataset_t *ds;
356 objset_t *mos = dp->dp_meta_objset;
357 hrtime_t start, write_time;
358 uint64_t data_written;
359 int err;
360 list_t synced_datasets;
361
362 list_create(&synced_datasets, sizeof (dsl_dataset_t),
363 offsetof(dsl_dataset_t, ds_synced_link));
364
365 /*
366 * We need to copy dp_space_towrite() before doing
367 * dsl_sync_task_sync(), because
368 * dsl_dataset_snapshot_reserve_space() will increase
369 * dp_space_towrite but not actually write anything.
370 */
371 data_written = dp->dp_space_towrite[txg & TXG_MASK];
372
373 tx = dmu_tx_create_assigned(dp, txg);
374
375 dp->dp_read_overhead = 0;
376 start = gethrtime();
377
378 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
379 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
380 /*
381 * We must not sync any non-MOS datasets twice, because
382 * we may have taken a snapshot of them. However, we
383 * may sync newly-created datasets on pass 2.
384 */
385 ASSERT(!list_link_active(&ds->ds_synced_link));
386 list_insert_tail(&synced_datasets, ds);
387 dsl_dataset_sync(ds, zio, tx);
388 }
389 DTRACE_PROBE(pool_sync__1setup);
390 err = zio_wait(zio);
391
392 write_time = gethrtime() - start;
393 ASSERT(err == 0);
394 DTRACE_PROBE(pool_sync__2rootzio);
395
396 /*
397 * After the data blocks have been written (ensured by the zio_wait()
398 * above), update the user/group space accounting.
399 */
400 for (ds = list_head(&synced_datasets); ds;
401 ds = list_next(&synced_datasets, ds))
402 dmu_objset_do_userquota_updates(ds->ds_objset, tx);
403
404 /*
405 * Sync the datasets again to push out the changes due to
406 * userspace updates. This must be done before we process the
407 * sync tasks, so that any snapshots will have the correct
408 * user accounting information (and we won't get confused
409 * about which blocks are part of the snapshot).
410 */
411 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
412 while (ds = txg_list_remove(&dp->dp_dirty_datasets, txg)) {
413 ASSERT(list_link_active(&ds->ds_synced_link));
414 dmu_buf_rele(ds->ds_dbuf, ds);
415 dsl_dataset_sync(ds, zio, tx);
416 }
417 err = zio_wait(zio);
418
419 /*
420 * Now that the datasets have been completely synced, we can
421 * clean up our in-memory structures accumulated while syncing:
422 *
423 * - move dead blocks from the pending deadlist to the on-disk deadlist
424 * - release hold from dsl_dataset_dirty()
425 */
426 while (ds = list_remove_head(&synced_datasets)) {
427 objset_t *os = ds->ds_objset;
428 bplist_iterate(&ds->ds_pending_deadlist,
429 deadlist_enqueue_cb, &ds->ds_deadlist, tx);
430 ASSERT(!dmu_objset_is_dirty(os, txg));
431 dmu_buf_rele(ds->ds_dbuf, ds);
432 }
433
434 start = gethrtime();
435 while (dd = txg_list_remove(&dp->dp_dirty_dirs, txg))
436 dsl_dir_sync(dd, tx);
437 write_time += gethrtime() - start;
438
439 /*
440 * The MOS's space is accounted for in the pool/$MOS
441 * (dp_mos_dir). We can't modify the mos while we're syncing
442 * it, so we remember the deltas and apply them here.
443 */
444 if (dp->dp_mos_used_delta != 0 || dp->dp_mos_compressed_delta != 0 ||
445 dp->dp_mos_uncompressed_delta != 0) {
446 dsl_dir_diduse_space(dp->dp_mos_dir, DD_USED_HEAD,
447 dp->dp_mos_used_delta,
448 dp->dp_mos_compressed_delta,
449 dp->dp_mos_uncompressed_delta, tx);
450 dp->dp_mos_used_delta = 0;
451 dp->dp_mos_compressed_delta = 0;
452 dp->dp_mos_uncompressed_delta = 0;
453 }
454
455 start = gethrtime();
456 if (list_head(&mos->os_dirty_dnodes[txg & TXG_MASK]) != NULL ||
457 list_head(&mos->os_free_dnodes[txg & TXG_MASK]) != NULL) {
458 zio = zio_root(dp->dp_spa, NULL, NULL, ZIO_FLAG_MUSTSUCCEED);
459 dmu_objset_sync(mos, zio, tx);
460 err = zio_wait(zio);
461 ASSERT(err == 0);
462 dprintf_bp(&dp->dp_meta_rootbp, "meta objset rootbp is %s", "");
463 spa_set_rootblkptr(dp->dp_spa, &dp->dp_meta_rootbp);
464 }
465 write_time += gethrtime() - start;
466 DTRACE_PROBE2(pool_sync__4io, hrtime_t, write_time,
467 hrtime_t, dp->dp_read_overhead);
468 write_time -= dp->dp_read_overhead;
469
470 /*
471 * If we modify a dataset in the same txg that we want to destroy it,
472 * its dsl_dir's dd_dbuf will be dirty, and thus have a hold on it.
473 * dsl_dir_destroy_check() will fail if there are unexpected holds.
474 * Therefore, we want to sync the MOS (thus syncing the dd_dbuf
475 * and clearing the hold on it) before we process the sync_tasks.
476 * The MOS data dirtied by the sync_tasks will be synced on the next
477 * pass.
478 */
479 DTRACE_PROBE(pool_sync__3task);
480 if (!txg_list_empty(&dp->dp_sync_tasks, txg)) {
481 dsl_sync_task_t *dst;
482 /*
483 * No more sync tasks should have been added while we
484 * were syncing.
485 */
486 ASSERT(spa_sync_pass(dp->dp_spa) == 1);
487 while (dst = txg_list_remove(&dp->dp_sync_tasks, txg))
488 dsl_sync_task_sync(dst, tx);
489 }
490
491 dmu_tx_commit(tx);
492
493 dp->dp_space_towrite[txg & TXG_MASK] = 0;
494 ASSERT(dp->dp_tempreserved[txg & TXG_MASK] == 0);
495
496 /*
497 * If the write limit max has not been explicitly set, set it
498 * to a fraction of available physical memory (default 1/8th).
499 * Note that we must inflate the limit because the spa
500 * inflates write sizes to account for data replication.
501 * Check this each sync phase to catch changing memory size.
502 */
503 if (physmem != old_physmem && zfs_write_limit_shift) {
504 mutex_enter(&zfs_write_limit_lock);
505 old_physmem = physmem;
506 zfs_write_limit_max = ptob(physmem) >> zfs_write_limit_shift;
507 zfs_write_limit_inflated = MAX(zfs_write_limit_min,
508 spa_get_asize(dp->dp_spa, zfs_write_limit_max));
509 mutex_exit(&zfs_write_limit_lock);
510 }
511
512 /*
513 * Attempt to keep the sync time consistent by adjusting the
514 * amount of write traffic allowed into each transaction group.
515 * Weight the throughput calculation towards the current value:
516 * thru = 3/4 old_thru + 1/4 new_thru
517 *
518 * Note: write_time is in nanosecs while dp_throughput is expressed in
519 * bytes per millisecond.
520 */
521 ASSERT(zfs_write_limit_min > 0);
522 if (data_written > zfs_write_limit_min / 8 &&
523 write_time > MSEC2NSEC(1)) {
524 uint64_t throughput = data_written / NSEC2MSEC(write_time);
525
526 if (dp->dp_throughput)
527 dp->dp_throughput = throughput / 4 +
528 3 * dp->dp_throughput / 4;
529 else
530 dp->dp_throughput = throughput;
531 dp->dp_write_limit = MIN(zfs_write_limit_inflated,
532 MAX(zfs_write_limit_min,
533 dp->dp_throughput * zfs_txg_synctime_ms));
534 }
535 }
536
537 void
538 dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
539 {
540 zilog_t *zilog;
541 dsl_dataset_t *ds;
542
543 while (zilog = txg_list_remove(&dp->dp_dirty_zilogs, txg)) {
544 ds = dmu_objset_ds(zilog->zl_os);
545 zil_clean(zilog, txg);
546 ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
547 dmu_buf_rele(ds->ds_dbuf, zilog);
548 }
549 ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
550 }
551
552 /*
553 * TRUE if the current thread is the tx_sync_thread or if we
554 * are being called from SPA context during pool initialization.
555 */
556 int
557 dsl_pool_sync_context(dsl_pool_t *dp)
558 {
559 return (curthread == dp->dp_tx.tx_sync_thread ||
560 spa_is_initializing(dp->dp_spa));
561 }
562
563 uint64_t
564 dsl_pool_adjustedsize(dsl_pool_t *dp, boolean_t netfree)
565 {
566 uint64_t space, resv;
567
568 /*
569 * Reserve about 1.6% (1/64), or at least 32MB, for allocation
570 * efficiency.
571 * XXX The intent log is not accounted for, so it must fit
572 * within this slop.
573 *
574 * If we're trying to assess whether it's OK to do a free,
575 * cut the reservation in half to allow forward progress
576 * (e.g. make it possible to rm(1) files from a full pool).
577 */
578 space = spa_get_dspace(dp->dp_spa);
579 resv = MAX(space >> 6, SPA_MINDEVSIZE >> 1);
580 if (netfree)
581 resv >>= 1;
582
583 return (space - resv);
584 }
585
586 int
587 dsl_pool_tempreserve_space(dsl_pool_t *dp, uint64_t space, dmu_tx_t *tx)
588 {
589 uint64_t reserved = 0;
590 uint64_t write_limit = (zfs_write_limit_override ?
591 zfs_write_limit_override : dp->dp_write_limit);
592
593 if (zfs_no_write_throttle) {
594 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK],
595 space);
596 return (0);
597 }
598
599 /*
600 * Check to see if we have exceeded the maximum allowed IO for
601 * this transaction group. We can do this without locks since
602 * a little slop here is ok. Note that we do the reserved check
603 * with only half the requested reserve: this is because the
604 * reserve requests are worst-case, and we really don't want to
605 * throttle based off of worst-case estimates.
606 */
607 if (write_limit > 0) {
608 reserved = dp->dp_space_towrite[tx->tx_txg & TXG_MASK]
609 + dp->dp_tempreserved[tx->tx_txg & TXG_MASK] / 2;
610
611 if (reserved && reserved > write_limit)
612 return (SET_ERROR(ERESTART));
613 }
614
615 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], space);
616
617 /*
618 * If this transaction group is over 7/8ths capacity, delay
619 * the caller 1 clock tick. This will slow down the "fill"
620 * rate until the sync process can catch up with us.
621 */
622 if (reserved && reserved > (write_limit - (write_limit >> 3))) {
623 txg_delay(dp, tx->tx_txg, zfs_throttle_delay,
624 zfs_throttle_resolution);
625 }
626
627 return (0);
628 }
629
630 void
631 dsl_pool_tempreserve_clear(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
632 {
633 ASSERT(dp->dp_tempreserved[tx->tx_txg & TXG_MASK] >= space);
634 atomic_add_64(&dp->dp_tempreserved[tx->tx_txg & TXG_MASK], -space);
635 }
636
637 void
638 dsl_pool_memory_pressure(dsl_pool_t *dp)
639 {
640 uint64_t space_inuse = 0;
641 int i;
642
643 if (dp->dp_write_limit == zfs_write_limit_min)
644 return;
645
646 for (i = 0; i < TXG_SIZE; i++) {
647 space_inuse += dp->dp_space_towrite[i];
648 space_inuse += dp->dp_tempreserved[i];
649 }
650 dp->dp_write_limit = MAX(zfs_write_limit_min,
651 MIN(dp->dp_write_limit, space_inuse / 4));
652 }
653
654 void
655 dsl_pool_willuse_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx)
656 {
657 if (space > 0) {
658 mutex_enter(&dp->dp_lock);
659 dp->dp_space_towrite[tx->tx_txg & TXG_MASK] += space;
660 mutex_exit(&dp->dp_lock);
661 }
662 }
663
664 /* ARGSUSED */
665 static int
666 upgrade_clones_cb(dsl_pool_t *dp, dsl_dataset_t *hds, void *arg)
667 {
668 dmu_tx_t *tx = arg;
669 dsl_dataset_t *ds, *prev = NULL;
670 int err;
671
672 err = dsl_dataset_hold_obj(dp, hds->ds_object, FTAG, &ds);
673 if (err)
674 return (err);
675
676 while (ds->ds_phys->ds_prev_snap_obj != 0) {
677 err = dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
678 FTAG, &prev);
679 if (err) {
680 dsl_dataset_rele(ds, FTAG);
681 return (err);
682 }
683
684 if (prev->ds_phys->ds_next_snap_obj != ds->ds_object)
685 break;
686 dsl_dataset_rele(ds, FTAG);
687 ds = prev;
688 prev = NULL;
689 }
690
691 if (prev == NULL) {
692 prev = dp->dp_origin_snap;
693
694 /*
695 * The $ORIGIN can't have any data, or the accounting
696 * will be wrong.
697 */
698 ASSERT0(prev->ds_phys->ds_bp.blk_birth);
699
700 /* The origin doesn't get attached to itself */
701 if (ds->ds_object == prev->ds_object) {
702 dsl_dataset_rele(ds, FTAG);
703 return (0);
704 }
705
706 dmu_buf_will_dirty(ds->ds_dbuf, tx);
707 ds->ds_phys->ds_prev_snap_obj = prev->ds_object;
708 ds->ds_phys->ds_prev_snap_txg = prev->ds_phys->ds_creation_txg;
709
710 dmu_buf_will_dirty(ds->ds_dir->dd_dbuf, tx);
711 ds->ds_dir->dd_phys->dd_origin_obj = prev->ds_object;
712
713 dmu_buf_will_dirty(prev->ds_dbuf, tx);
714 prev->ds_phys->ds_num_children++;
715
716 if (ds->ds_phys->ds_next_snap_obj == 0) {
717 ASSERT(ds->ds_prev == NULL);
718 VERIFY0(dsl_dataset_hold_obj(dp,
719 ds->ds_phys->ds_prev_snap_obj, ds, &ds->ds_prev));
720 }
721 }
722
723 ASSERT3U(ds->ds_dir->dd_phys->dd_origin_obj, ==, prev->ds_object);
724 ASSERT3U(ds->ds_phys->ds_prev_snap_obj, ==, prev->ds_object);
725
726 if (prev->ds_phys->ds_next_clones_obj == 0) {
727 dmu_buf_will_dirty(prev->ds_dbuf, tx);
728 prev->ds_phys->ds_next_clones_obj =
729 zap_create(dp->dp_meta_objset,
730 DMU_OT_NEXT_CLONES, DMU_OT_NONE, 0, tx);
731 }
732 VERIFY0(zap_add_int(dp->dp_meta_objset,
733 prev->ds_phys->ds_next_clones_obj, ds->ds_object, tx));
734
735 dsl_dataset_rele(ds, FTAG);
736 if (prev != dp->dp_origin_snap)
737 dsl_dataset_rele(prev, FTAG);
738 return (0);
739 }
740
741 void
742 dsl_pool_upgrade_clones(dsl_pool_t *dp, dmu_tx_t *tx)
743 {
744 ASSERT(dmu_tx_is_syncing(tx));
745 ASSERT(dp->dp_origin_snap != NULL);
746
747 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj, upgrade_clones_cb,
748 tx, DS_FIND_CHILDREN));
749 }
750
751 /* ARGSUSED */
752 static int
753 upgrade_dir_clones_cb(dsl_pool_t *dp, dsl_dataset_t *ds, void *arg)
754 {
755 dmu_tx_t *tx = arg;
756 objset_t *mos = dp->dp_meta_objset;
757
758 if (ds->ds_dir->dd_phys->dd_origin_obj != 0) {
759 dsl_dataset_t *origin;
760
761 VERIFY0(dsl_dataset_hold_obj(dp,
762 ds->ds_dir->dd_phys->dd_origin_obj, FTAG, &origin));
763
764 if (origin->ds_dir->dd_phys->dd_clones == 0) {
765 dmu_buf_will_dirty(origin->ds_dir->dd_dbuf, tx);
766 origin->ds_dir->dd_phys->dd_clones = zap_create(mos,
767 DMU_OT_DSL_CLONES, DMU_OT_NONE, 0, tx);
768 }
769
770 VERIFY0(zap_add_int(dp->dp_meta_objset,
771 origin->ds_dir->dd_phys->dd_clones, ds->ds_object, tx));
772
773 dsl_dataset_rele(origin, FTAG);
774 }
775 return (0);
776 }
777
778 void
779 dsl_pool_upgrade_dir_clones(dsl_pool_t *dp, dmu_tx_t *tx)
780 {
781 ASSERT(dmu_tx_is_syncing(tx));
782 uint64_t obj;
783
784 (void) dsl_dir_create_sync(dp, dp->dp_root_dir, FREE_DIR_NAME, tx);
785 VERIFY0(dsl_pool_open_special_dir(dp,
786 FREE_DIR_NAME, &dp->dp_free_dir));
787
788 /*
789 * We can't use bpobj_alloc(), because spa_version() still
790 * returns the old version, and we need a new-version bpobj with
791 * subobj support. So call dmu_object_alloc() directly.
792 */
793 obj = dmu_object_alloc(dp->dp_meta_objset, DMU_OT_BPOBJ,
794 SPA_MAXBLOCKSIZE, DMU_OT_BPOBJ_HDR, sizeof (bpobj_phys_t), tx);
795 VERIFY0(zap_add(dp->dp_meta_objset, DMU_POOL_DIRECTORY_OBJECT,
796 DMU_POOL_FREE_BPOBJ, sizeof (uint64_t), 1, &obj, tx));
797 VERIFY0(bpobj_open(&dp->dp_free_bpobj, dp->dp_meta_objset, obj));
798
799 VERIFY0(dmu_objset_find_dp(dp, dp->dp_root_dir_obj,
800 upgrade_dir_clones_cb, tx, DS_FIND_CHILDREN));
801 }
802
803 void
804 dsl_pool_create_origin(dsl_pool_t *dp, dmu_tx_t *tx)
805 {
806 uint64_t dsobj;
807 dsl_dataset_t *ds;
808
809 ASSERT(dmu_tx_is_syncing(tx));
810 ASSERT(dp->dp_origin_snap == NULL);
811 ASSERT(rrw_held(&dp->dp_config_rwlock, RW_WRITER));
812
813 /* create the origin dir, ds, & snap-ds */
814 dsobj = dsl_dataset_create_sync(dp->dp_root_dir, ORIGIN_DIR_NAME,
815 NULL, 0, kcred, tx);
816 VERIFY0(dsl_dataset_hold_obj(dp, dsobj, FTAG, &ds));
817 dsl_dataset_snapshot_sync_impl(ds, ORIGIN_DIR_NAME, tx);
818 VERIFY0(dsl_dataset_hold_obj(dp, ds->ds_phys->ds_prev_snap_obj,
819 dp, &dp->dp_origin_snap));
820 dsl_dataset_rele(ds, FTAG);
821 }
822
823 taskq_t *
824 dsl_pool_vnrele_taskq(dsl_pool_t *dp)
825 {
826 return (dp->dp_vnrele_taskq);
827 }
828
829 /*
830 * Walk through the pool-wide zap object of temporary snapshot user holds
831 * and release them.
832 */
833 void
834 dsl_pool_clean_tmp_userrefs(dsl_pool_t *dp)
835 {
836 zap_attribute_t za;
837 zap_cursor_t zc;
838 objset_t *mos = dp->dp_meta_objset;
839 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
840 nvlist_t *holds;
841
842 if (zapobj == 0)
843 return;
844 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
845
846 holds = fnvlist_alloc();
847
848 for (zap_cursor_init(&zc, mos, zapobj);
849 zap_cursor_retrieve(&zc, &za) == 0;
850 zap_cursor_advance(&zc)) {
851 char *htag;
852 uint64_t dsobj;
853 nvlist_t *tags;
854
855 htag = strchr(za.za_name, '-');
856 *htag = '\0';
857 ++htag;
858 if (nvlist_lookup_nvlist(holds, za.za_name, &tags) != 0) {
859 tags = fnvlist_alloc();
860 fnvlist_add_boolean(tags, htag);
861 fnvlist_add_nvlist(holds, za.za_name, tags);
862 fnvlist_free(tags);
863 } else {
864 fnvlist_add_boolean(tags, htag);
865 }
866 }
867 dsl_dataset_user_release_tmp(dp, holds);
868 fnvlist_free(holds);
869 zap_cursor_fini(&zc);
870 }
871
872 /*
873 * Create the pool-wide zap object for storing temporary snapshot holds.
874 */
875 void
876 dsl_pool_user_hold_create_obj(dsl_pool_t *dp, dmu_tx_t *tx)
877 {
878 objset_t *mos = dp->dp_meta_objset;
879
880 ASSERT(dp->dp_tmp_userrefs_obj == 0);
881 ASSERT(dmu_tx_is_syncing(tx));
882
883 dp->dp_tmp_userrefs_obj = zap_create_link(mos, DMU_OT_USERREFS,
884 DMU_POOL_DIRECTORY_OBJECT, DMU_POOL_TMP_USERREFS, tx);
885 }
886
887 static int
888 dsl_pool_user_hold_rele_impl(dsl_pool_t *dp, uint64_t dsobj,
889 const char *tag, uint64_t now, dmu_tx_t *tx, boolean_t holding)
890 {
891 objset_t *mos = dp->dp_meta_objset;
892 uint64_t zapobj = dp->dp_tmp_userrefs_obj;
893 char *name;
894 int error;
895
896 ASSERT(spa_version(dp->dp_spa) >= SPA_VERSION_USERREFS);
897 ASSERT(dmu_tx_is_syncing(tx));
898
899 /*
900 * If the pool was created prior to SPA_VERSION_USERREFS, the
901 * zap object for temporary holds might not exist yet.
902 */
903 if (zapobj == 0) {
904 if (holding) {
905 dsl_pool_user_hold_create_obj(dp, tx);
906 zapobj = dp->dp_tmp_userrefs_obj;
907 } else {
908 return (SET_ERROR(ENOENT));
909 }
910 }
911
912 name = kmem_asprintf("%llx-%s", (u_longlong_t)dsobj, tag);
913 if (holding)
914 error = zap_add(mos, zapobj, name, 8, 1, &now, tx);
915 else
916 error = zap_remove(mos, zapobj, name, tx);
917 strfree(name);
918
919 return (error);
920 }
921
922 /*
923 * Add a temporary hold for the given dataset object and tag.
924 */
925 int
926 dsl_pool_user_hold(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
927 uint64_t now, dmu_tx_t *tx)
928 {
929 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, now, tx, B_TRUE));
930 }
931
932 /*
933 * Release a temporary hold for the given dataset object and tag.
934 */
935 int
936 dsl_pool_user_release(dsl_pool_t *dp, uint64_t dsobj, const char *tag,
937 dmu_tx_t *tx)
938 {
939 return (dsl_pool_user_hold_rele_impl(dp, dsobj, tag, NULL,
940 tx, B_FALSE));
941 }
942
943 /*
944 * DSL Pool Configuration Lock
945 *
946 * The dp_config_rwlock protects against changes to DSL state (e.g. dataset
947 * creation / destruction / rename / property setting). It must be held for
948 * read to hold a dataset or dsl_dir. I.e. you must call
949 * dsl_pool_config_enter() or dsl_pool_hold() before calling
950 * dsl_{dataset,dir}_hold{_obj}. In most circumstances, the dp_config_rwlock
951 * must be held continuously until all datasets and dsl_dirs are released.
952 *
953 * The only exception to this rule is that if a "long hold" is placed on
954 * a dataset, then the dp_config_rwlock may be dropped while the dataset
955 * is still held. The long hold will prevent the dataset from being
956 * destroyed -- the destroy will fail with EBUSY. A long hold can be
957 * obtained by calling dsl_dataset_long_hold(), or by "owning" a dataset
958 * (by calling dsl_{dataset,objset}_{try}own{_obj}).
959 *
960 * Legitimate long-holders (including owners) should be long-running, cancelable
961 * tasks that should cause "zfs destroy" to fail. This includes DMU
962 * consumers (i.e. a ZPL filesystem being mounted or ZVOL being open),
963 * "zfs send", and "zfs diff". There are several other long-holders whose
964 * uses are suboptimal (e.g. "zfs promote", and zil_suspend()).
965 *
966 * The usual formula for long-holding would be:
967 * dsl_pool_hold()
968 * dsl_dataset_hold()
969 * ... perform checks ...
970 * dsl_dataset_long_hold()
971 * dsl_pool_rele()
972 * ... perform long-running task ...
973 * dsl_dataset_long_rele()
974 * dsl_dataset_rele()
975 *
976 * Note that when the long hold is released, the dataset is still held but
977 * the pool is not held. The dataset may change arbitrarily during this time
978 * (e.g. it could be destroyed). Therefore you shouldn't do anything to the
979 * dataset except release it.
980 *
981 * User-initiated operations (e.g. ioctls, zfs_ioc_*()) are either read-only
982 * or modifying operations.
983 *
984 * Modifying operations should generally use dsl_sync_task(). The synctask
985 * infrastructure enforces proper locking strategy with respect to the
986 * dp_config_rwlock. See the comment above dsl_sync_task() for details.
987 *
988 * Read-only operations will manually hold the pool, then the dataset, obtain
989 * information from the dataset, then release the pool and dataset.
990 * dmu_objset_{hold,rele}() are convenience routines that also do the pool
991 * hold/rele.
992 */
993
994 int
995 dsl_pool_hold(const char *name, void *tag, dsl_pool_t **dp)
996 {
997 spa_t *spa;
998 int error;
999
1000 error = spa_open(name, &spa, tag);
1001 if (error == 0) {
1002 *dp = spa_get_dsl(spa);
1003 dsl_pool_config_enter(*dp, tag);
1004 }
1005 return (error);
1006 }
1007
1008 void
1009 dsl_pool_rele(dsl_pool_t *dp, void *tag)
1010 {
1011 dsl_pool_config_exit(dp, tag);
1012 spa_close(dp->dp_spa, tag);
1013 }
1014
1015 void
1016 dsl_pool_config_enter(dsl_pool_t *dp, void *tag)
1017 {
1018 /*
1019 * We use a "reentrant" reader-writer lock, but not reentrantly.
1020 *
1021 * The rrwlock can (with the track_all flag) track all reading threads,
1022 * which is very useful for debugging which code path failed to release
1023 * the lock, and for verifying that the *current* thread does hold
1024 * the lock.
1025 *
1026 * (Unlike a rwlock, which knows that N threads hold it for
1027 * read, but not *which* threads, so rw_held(RW_READER) returns TRUE
1028 * if any thread holds it for read, even if this thread doesn't).
1029 */
1030 ASSERT(!rrw_held(&dp->dp_config_rwlock, RW_READER));
1031 rrw_enter(&dp->dp_config_rwlock, RW_READER, tag);
1032 }
1033
1034 void
1035 dsl_pool_config_exit(dsl_pool_t *dp, void *tag)
1036 {
1037 rrw_exit(&dp->dp_config_rwlock, tag);
1038 }
1039
1040 boolean_t
1041 dsl_pool_config_held(dsl_pool_t *dp)
1042 {
1043 return (RRW_LOCK_HELD(&dp->dp_config_rwlock));
1044 }